Add Assistent reference books and remove training dataset output.

Introduce books/ with civitai-krea2 and HF fictext builders, sha-diff seed to VM, and drop train.jsonl from the civitai scrape pipeline. Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-23 22:56:07 +03:00
parent 84ed0bb47a
commit 07bb521b70
27 changed files with 836 additions and 64 deletions
@@ -29,13 +29,18 @@ Output/**
 !Output/**/.gitkeep
 !Output/README.md

-# Civitai Krea2 scrape (meta only — train.jsonl / search.jsonl)
+# Civitai Krea2 scrape + Assistent books (gitignore data)
 datasets/**
 !datasets/.gitkeep
 !datasets/README.md
 !datasets/**/.gitkeep
 !datasets/**/README.md

+books/**
+!books/.gitkeep
+!books/README.md
+!books.example/**
+
 # Python
 .venv/
 __pycache__/
@@ -0,0 +1,4 @@
+@echo off
+setlocal
+cd /d "%~dp0"
+python -m gpu_rent.books %*
@@ -0,0 +1,2 @@
+Set-Location $PSScriptRoot
+python -m gpu_rent.books @args
@@ -0,0 +1,4 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd "$(dirname "$0")"
+exec python -m gpu_rent.books "$@"
@@ -0,0 +1,14 @@
+kind: book
+id: civitai-krea2
+title: Civitai Krea2
+description: >-
+  Примеры промптов, негативов, steps/cfg и LoRA для Krea 2 с Civitai.
+  Ищи сюда, когда нужен рабочий prompt или параметры генерации.
+content_kind: prompt-examples
+language: en
+tags:
+  - civitai
+  - krea2
+  - prompts
+source: civitai-dataset
+license: civitai-terms
@@ -0,0 +1,27 @@
+# Assistent reference books (local)
+
+FTS `search.jsonl` indexes for Swarm Assistent Knowledge Hub. **Not** in Typer main CLI — use:
+
+```text
+books-dataset.bat              # python -m gpu_rent.books
+python -m gpu_rent.books list
+python -m gpu_rent.books build civitai
+python -m gpu_rent.books download fictext --variant both
+gpu-rent seed-books            # push changed books → VM
+```
+
+## Layout
+
+| Path | Role |
+| --- | --- |
+| `civitai-krea2/` | From `civitai-dataset split` → `book.yaml` + `search.jsonl` + `meta.json` |
+| `ru-fictext-rplus/` | HF `ru-fictext-nsfw-data-r+.parquet` → chunked search rows |
+| `ru-fictext-full/` | HF full parquet |
+
+On `up` / `seed-personas` / `seed-books`: changed books (by `meta.content_sha`) → `/mnt/swarm_data/Assistent/books/<id>/`.
+
+Civitai scrape: manual on laptop; `up` pushes when local book changed.
+
+Fictext: `pip install 'gpu-rent[books]'` (pyarrow) for build.
+
+See [docs/local-folders.md](../docs/local-folders.md), [datasets/README.md](../datasets/README.md).
@@ -1,13 +1,15 @@
 # Civitai Krea2 datasets (local)

-Scraped generation metadata for Assistent training + cheap FTS search. **No image files.**
+Scraped generation metadata for Assistent **books** (FTS search). **No image files.** UI training (`train.jsonl`) removed.

-Not part of `gpu-rent` Typer CLI and **not** `models.yaml` / `gpu-rent push`. Separate entrypoint:
+Not part of `gpu-rent` Typer CLI and **not** `models.yaml` / `gpu-rent push`. Separate entrypoints:

 ```text
 civitai-dataset.bat          # Windows (or .\civitai-dataset.ps1 / ./civitai-dataset.sh)
-# default = all → discover + scrape + split
+# default = all → discover + scrape + split (+ books/civitai-krea2)
 python -m gpu_rent.civitai_dataset all
+
+books-dataset.bat            # python -m gpu_rent.books download fictext …
 ```

 Needs `CIVITAI_API_TOKEN` and preferably `CIVITAI_API_HOST=civitai.red` in `.env`.
@@ -18,7 +20,7 @@ Needs `CIVITAI_API_TOKEN` and preferably `CIVITAI_API_HOST=civitai.red` in `.env
 | --- | --- | --- |
 | `discover` | `--top-checkpoints` (20), `--top-loras` (30) | Popular Krea2 ckpt/LoRAs (+ versions from `models.yaml`) → `catalog/models.json` |
 | `scrape` | `--target` (2000), `--per-version`, `--min-score` | Gallery metadata only; resume by id; NSFW→rating; skip minor tags |
-| `split` | — | `by_kind/` / `by_rating/` / `train.jsonl` / `search.jsonl` |
+| `split` | — | `by_kind/` / `by_rating/` / `search.jsonl` + `books/civitai-krea2/` |
 | `all` | same as discover+scrape | discover → scrape → split (launcher default) |

 Global: `--out` = app root (writes under `datasets/civitai`).
@@ -31,11 +33,9 @@ Global: `--out` = app root (writes under `datasets/civitai`).
 | `civitai/catalog/images.jsonl` | Canonical rows (resume by id) |
 | `civitai/by_kind/` | checkpoint vs lora |
 | `civitai/by_rating/` | pg / pg13 / r / x |
-| `civitai/train.jsonl` | Alpaca instruction/output for Assistent Обучение |
-| `civitai/search.jsonl` | Compact FTS index → pushed to VM as `Assistent/civitai-examples.jsonl` on `up` / `seed-personas` |
+| `civitai/search.jsonl` | Compact Civitai FTS source → `books/civitai-krea2` on split |
+| `../books/civitai-krea2/` | Assistent book (`book.yaml`, `search.jsonl`, `meta.json`) |

-Push of `search.jsonl` is skipped if the file is missing. Weights stay in [models.md](../docs/models.md) (Civitai seed / `Models/` / `capture`).
+On `up` / `seed-personas` / `seed-books`: changed books push → `/mnt/swarm_data/Assistent/books/` (sha diff). Legacy `civitai-examples.jsonl` only if `books/` empty.

-Git ignores the data; this README stays.
-
-See also: [docs/cli.md](../docs/cli.md) (civitai-dataset block), [docs/extensions.md](../docs/extensions.md), [docs/local-folders.md](../docs/local-folders.md).
+See [books/README.md](../books/README.md), [docs/cli.md](../docs/cli.md), [docs/local-folders.md](../docs/local-folders.md).
@@ -44,8 +44,9 @@
 | `inventory` / `ux` / `placement` / `flavor_presets` | Flavors, квоты, AZ; preview / ServerPlan |
 | `session` | `cmd_up` / `cmd_stop` / adopt |
 | `bootstrap` + `remote/bootstrap.sh` | Идемпотентный first-boot; light без apt |
-| `provision` | extensions, autocomplete, civitai seed, LLM, push, idle-killer arm, start SwarmUI; seed `search.jsonl` → `Assistent/civitai-examples.jsonl` |
-| `civitai` / `civitai_dataset` | Civitai HTTP API; локальный scrape Krea2 → train/search jsonl (отдельный CLI) |
+| `provision` | extensions, autocomplete, books seed, LLM, push, idle-killer arm, start SwarmUI; `seed_books()` → `Assistent/books/` (sha diff) |
+| `civitai` / `civitai_dataset` | Civitai HTTP API; локальный scrape Krea2 → search jsonl + book `books/civitai-krea2` |
+| `books` | HF fictext + civitai book build; `meta.content_sha`; отдельный CLI + `seed-books` |
 | `huggingface` | HF probe / метадата / URL для seed и capture |
 | `capture` | Инвентарь VM → merge ссылок в локальные манифесты |
 | `doctor` | Preflight без mutating compute |
@@ -79,7 +79,8 @@ gpu-rent up --yes --ollama
 | `gpu-rent ssh` | Оболочка на VM |
 | `gpu-rent seed-models` | Докачать новые строки Civitai-манифеста на живой диск |
 | `gpu-rent seed-extensions` | Доклонировать/обновить git-репы; restart swarmui |
-| `gpu-rent seed-personas` | Пуш `assistent-extensions/` → `Assistent/extensions/` (+ `_base`); также `search.jsonl` → FTS, если есть |
+| `gpu-rent seed-personas` | Пуш `assistent-extensions/` → `Assistent/extensions/` (+ `_base`); также `seed-books` если есть `books/` |
+| `gpu-rent seed-books` | Push изменённых `books/*/` → `Assistent/books/` (`--force` — все) |
 | `gpu-rent push` / `push-models` | Локальные деревья → VM |
 | `gpu-rent pull-output` | VM `Output/` → `./Output` |
 | `gpu-rent capture` / `capture all` | Инвентарь VM → merge **ссылок** в `models.yaml` + `extensions.yaml` (веса не качать) |
@@ -102,10 +103,10 @@ gpu-rent up --yes --ollama
 | --- | --- |
 | `discover` | Топ Krea2 checkpoints/LoRAs → `catalog/models.json` |
 | `scrape` | Галереи (метаданные, **без** картинок); `--target`, `--per-version`, `--min-score` |
-| `split` | `by_kind/` / `by_rating/` / `train.jsonl` / `search.jsonl` |
+| `split` | `by_kind/` / `by_rating/` / `search.jsonl` + `books/civitai-krea2/` |
 | `all` | discover + scrape + split |

-Нужен `CIVITAI_API_TOKEN` (лучше `CIVITAI_API_HOST=civitai.red`). На `up` / `seed-personas` файл `datasets/civitai/search.jsonl` пушится на VM как `Assistent/civitai-examples.jsonl` (если есть).
+Нужен `CIVITAI_API_TOKEN` (лучше `CIVITAI_API_HOST=civitai.red`). Книги: `books-dataset.bat` / `python -m gpu_rent.books download fictext --variant both` (pyarrow: `pip install 'gpu-rent[books]'`). На `up` / `seed-books` изменённые `books/*/` → `Assistent/books/`.

 ---

@@ -62,4 +62,4 @@
 | LLM runtime | Только `none` \| `ollama`. **llama.cpp снят** из продукта; leftover unit `gpu-rent-llamacpp` на `up` стопается |
 | Debug API | Принятая поверхность: localhost `:17821` на `up`/`tunnel` / `gpu-rent debug`; Assistent diagnose, session, chat-eval |
 | Assistent | First-class поверх SwarmUI + Ollama (чат, Cards, personas), не только «prompt-help» |
-| Civitai dataset | Локальный scrape метаданных Krea2 (без картинок) → `train.jsonl` / `search.jsonl`; на `up`/`seed-personas` `search.jsonl` → `Assistent/civitai-examples.jsonl` (FTS) |
+| Civitai dataset | Локальный scrape Krea2 (без картинок) → `search.jsonl` + book `books/civitai-krea2/`; на `up`/`seed-books` push при изменении `meta.content_sha` → `Assistent/books/` (FTS). Legacy `civitai-examples.jsonl` если books нет |
@@ -70,7 +70,7 @@ Bundled в swarm-assistent: `neutral` (Нормальный), `aggressive`, `dre

 Формат пака: `assistent-pack.yaml` (`kind: persona`, `id: …`) + полки JSON в корне (или `personas/<id>/` для мульти-пака). На `up` также пишется `_base/assistant.json` (`default_persona: neutral`, опционально `num_ctx` с GPU tier). Seed **не** удаляет overlay-личности, созданные в UI. Legacy `assistent-personas/` / `personas.json` больше не сидятся.

-Если есть `datasets/civitai/search.jsonl`, на том же шаге пушится FTS-индекс → `/mnt/swarm_data/Assistent/civitai-examples.jsonl` ([datasets/README](../datasets/README.md)). Диагностика Assistent: Debug API `/assistent/*` ([cli.md](cli.md)).
+Если есть `books/` (или legacy `datasets/civitai/search.jsonl`), на том же шаге `seed-books` пушит FTS → `/mnt/swarm_data/Assistent/books/` ([books/README](../books/README.md), [knowledge-books.md](knowledge-books.md)). Диагностика Assistent: Debug API `/assistent/*` ([cli.md](cli.md)).

 В git репозитория gpu-rent не коммитить рабочий список с лишними приватными URL сверх примера.

@@ -0,0 +1,47 @@
+# Assistent Knowledge Hub (swarm-assistent)
+
+gpu-rent builds and seeds **books** locally. Runtime lives in [swarm-assistent](https://gitea.hsrv.site/mrleo1nid/swarm-assistent.git).
+
+## VM layout
+
+```text
+/mnt/swarm_data/Assistent/books/<book-id>/
+  book.yaml          # title, description, content_kind
+  search.jsonl       # FTS rows (unified contract)
+  meta.json          # row_count, content_sha, built_at
+  .gpu-rent-meta.json  # same sha for gpu-rent skip/push
+```
+
+## Unified search row
+
+```json
+{
+  "id": "civitai-krea2:42",
+  "book": "civitai-krea2",
+  "title": "",
+  "tags": ["woman"],
+  "text": "excerpt ≤500 chars",
+  "body": "longer chunk",
+  "rating": "pg13",
+  "meta": {}
+}
+```
+
+## swarm-assistent tasks (not in gpu-rent)
+
+1. **`Knowledge/IKnowledgeSource.cs`** + adapters: Book, MemorySeed, MemoryPersonal, Tag, Session
+2. **`AssistentKnowledgeHub`**: ListCatalog, Search, Upsert, Forget, RunPipeline, BuildSystemLayer
+3. **Remove training UI**: Обучение tab, AssistentTrainingJobs, AssistentListTrainSamples, taste APIs
+4. **HTTP** `AssistentChat` → `knowledge: {catalog, hops, results}`; legacy `civitai_results` shim
+5. **Persona UI** «Знания»: attach toggles; bundled personas default attach civitai-krea2 + ru-fictext-rplus
+6. **Skill `knowledge`**: list / search / upsert / forget
+
+## Default attach (bundled only)
+
+- `neutral`, `aggressive`, `dreamer`: civitai-krea2 + ru-fictext-rplus
+- UI-created personas: no auto-attach
+- Pack `leonid`: from `assistent-pack.yaml` `knowledge.attach`
+
+## Index on VM
+
+On book push or sha change: reindex `search.jsonl` → sqlite FTS `books_fts`. Progress in UI on first chat.
@@ -113,7 +113,7 @@ Unit `gpu-rent-ollama` читает `/mnt/swarm_data/.gpu-rent-gpu.json`:

 \*Flash на Ampere+ (compute ≥ 8.0). `NUM_PARALLEL=2`, `MAX_LOADED_MODELS=2` (chat VL + memory embed). Memory models use `use: memory` and a CPU Modelfile (`num_gpu 0`) so embed does not steal VRAM from the chat model. Ollama default `num_ctx` is 4096; we set `OLLAMA_CONTEXT_LENGTH` so Assistent + vision fits. Mid/high keep-alive is **5m** because Assistent parks the chat model before Generate (`AssistentParkLlm`) and warms after. Env: `/mnt/swarm_data/.gpu-rent-ollama.env`.

-Personas: bundled `neutral` / `aggressive` / `dreamer` in swarm-assistent; extra packs via `extensions.yaml` `assistent:` or laptop `assistent-extensions/` → `Assistent/extensions/` on `up` / `seed-personas` / `push` (+ `_base/assistant.json` with `default_persona` / `num_ctx`). See [extensions.md](extensions.md). Optional FTS examples: [datasets/README](../datasets/README.md) (`search.jsonl` → `Assistent/civitai-examples.jsonl`). Journal: `gpu-rent logs -u ollama`.
+Personas: bundled `neutral` / `aggressive` / `dreamer` in swarm-assistent; extra packs via `extensions.yaml` `assistent:` or laptop `assistent-extensions/` → `Assistent/extensions/` on `up` / `seed-personas` / `push` (+ `_base/assistant.json` with `default_persona` / `num_ctx`). See [extensions.md](extensions.md). Reference books: [books/README](../books/README.md) → `Assistent/books/` on `seed-books` ([knowledge-books.md](knowledge-books.md)). Journal: `gpu-rent logs -u ollama`.

 `ollama-models.yaml` entries:

@@ -26,7 +26,8 @@
 | Локально | На VM | Когда |
 | --- | --- | --- |
 | `assistent-extensions/<pack>/` | `/mnt/swarm_data/Assistent/extensions/<pack>/` | `up` / `seed-personas` / `push` (без `.git`); см. [extensions.md](extensions.md) |
-| `datasets/civitai/search.jsonl` | `Assistent/civitai-examples.jsonl` | `up` / `seed-personas`, если файл есть; см. [datasets/README](../datasets/README.md) |
+| `datasets/civitai/search.jsonl` | (local build) → `books/civitai-krea2/` | civitai-dataset split |
+| `books/<id>/` | `Assistent/books/<id>/` | `up` / `seed-personas` / `seed-books` if `meta.content_sha` changed; see [books/README](../books/README.md) |

 ## Push (Models / Wildcards / CustomWorkflows / packs)

@@ -90,7 +90,7 @@
 ## Уже в коде после v1

 - Debug API на `127.0.0.1:17821` (`gpu-rent debug` / sidecar на `up`/`tunnel`); Assistent session / chat-eval
- `civitai-dataset` → `datasets/civitai/search.jsonl` → FTS push на `up` / `seed-personas`
+- `civitai-dataset` → `books/civitai-krea2/`; `gpu-rent books download fictext`; `seed-books` на `up` / `seed-personas` (sha diff)
 - `seed-personas`, `capture wanted`, `logs --unit swarm|ollama|killer|cloud-init`
 - Balance notify (`SELECTEL_API_TOKEN` + watchdog)

@@ -31,6 +31,7 @@ dependencies = [

 [project.optional-dependencies]
 dev = ["pytest>=8.0"]
+books = ["pyarrow>=15.0"]

 [project.scripts]
 gpu-rent = "gpu_rent.cli:app"
@@ -0,0 +1,458 @@
+"""Assistent reference books: build locally, seed to VM on change (FTS search.jsonl)."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import re
+import sys
+from collections.abc import Callable, Iterable
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+import httpx
+import yaml
+
+from gpu_rent.civitai_dataset import looks_minor, read_jsonl, write_jsonl
+from gpu_rent.config import load_config
+from gpu_rent.errors import GpuRentError
+from gpu_rent.paths import app_root
+
+Log = Callable[[str], None]
+
+BOOK_CIVITAI = "civitai-krea2"
+BOOK_FICTEXT_RPLUS = "ru-fictext-rplus"
+BOOK_FICTEXT_FULL = "ru-fictext-full"
+
+VM_BOOKS_PREFIX = "/mnt/swarm_data/Assistent/books"
+SIDEcar_NAME = ".gpu-rent-meta.json"
+
+HF_REPO = "krplt/ru-fictext-nsfw"
+HF_FILES = {
+    BOOK_FICTEXT_RPLUS: "ru-fictext-nsfw-data-r+.parquet",
+    BOOK_FICTEXT_FULL: "ru-fictext-nsfw-data.parquet",
+}
+
+CHUNK_CHARS = 500
+CHUNK_OVERLAP = 80
+EXCERPT_MAX = 500
+
+BOOK_SPECS: dict[str, dict[str, Any]] = {
+    BOOK_CIVITAI: {
+        "kind": "book",
+        "id": BOOK_CIVITAI,
+        "title": "Civitai Krea2",
+        "description": (
+            "Примеры промптов, негативов, steps/cfg и LoRA для Krea 2 с Civitai. "
+            "Ищи сюда, когда нужен рабочий prompt или параметры генерации."
+        ),
+        "content_kind": "prompt-examples",
+        "language": "en",
+        "tags": ["civitai", "krea2", "prompts"],
+        "source": "civitai-dataset",
+        "license": "civitai-terms",
+    },
+    BOOK_FICTEXT_RPLUS: {
+        "kind": "book",
+        "id": BOOK_FICTEXT_RPLUS,
+        "title": "RU ficbook R+",
+        "description": (
+            "Отрывки русской фанфикшн (mature/NSFW) для стиля, лексики и тона сцен. "
+            "Не цитируй дословно длинно — бери формулировки и ритм."
+        ),
+        "content_kind": "prose-style",
+        "language": "ru",
+        "tags": ["fanfiction", "nsfw", "russian"],
+        "source": f"huggingface.co/datasets/{HF_REPO}",
+        "license": "cc-by-4.0",
+    },
+    BOOK_FICTEXT_FULL: {
+        "kind": "book",
+        "id": BOOK_FICTEXT_FULL,
+        "title": "RU ficbook (full)",
+        "description": (
+            "Отрывки русской фанфикшн (NSFW + safe) для стиля и лексики. "
+            "Шире чем R+; для нейтральных сцен предпочитай safe-теги."
+        ),
+        "content_kind": "prose-style",
+        "language": "ru",
+        "tags": ["fanfiction", "russian"],
+        "source": f"huggingface.co/datasets/{HF_REPO}",
+        "license": "cc-by-4.0",
+    },
+}
+
+
+def _log(msg: str) -> None:
+    print(msg, flush=True)
+
+
+def books_root(root: Path | None = None) -> Path:
+    return (root or app_root()) / "books"
+
+
+def book_dir(book_id: str, root: Path | None = None) -> Path:
+    return books_root(root) / book_id
+
+
+def write_book_yaml(book_id: str, root: Path | None = None) -> Path:
+    spec = BOOK_SPECS.get(book_id)
+    if not spec:
+        raise GpuRentError(f"Unknown book id: {book_id}")
+    path = book_dir(book_id, root) / "book.yaml"
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(yaml.safe_dump(spec, allow_unicode=True, sort_keys=False), encoding="utf-8")
+    return path
+
+
+def sha256_file(path: Path) -> str:
+    h = hashlib.sha256()
+    with path.open("rb") as fh:
+        for chunk in iter(lambda: fh.read(1024 * 1024), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+def write_meta(
+    book_id: str,
+    *,
+    row_count: int,
+    root: Path | None = None,
+    source_sha: str | None = None,
+    extra: dict[str, Any] | None = None,
+) -> Path:
+    bdir = book_dir(book_id, root)
+    search = bdir / "search.jsonl"
+    if not search.is_file():
+        raise GpuRentError(f"Нет {search} для meta")
+    payload: dict[str, Any] = {
+        "book_id": book_id,
+        "row_count": row_count,
+        "content_sha": sha256_file(search),
+        "built_at": datetime.now(UTC).replace(microsecond=0).isoformat().replace("+00:00", "Z"),
+        "search_bytes": search.stat().st_size,
+    }
+    if source_sha:
+        payload["source_sha"] = source_sha
+    if extra:
+        payload.update(extra)
+    path = bdir / "meta.json"
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+    return path
+
+
+def read_meta(book_id: str, root: Path | None = None) -> dict[str, Any] | None:
+    path = book_dir(book_id, root) / "meta.json"
+    if not path.is_file():
+        return None
+    try:
+        data = json.loads(path.read_text(encoding="utf-8"))
+    except json.JSONDecodeError:
+        return None
+    return data if isinstance(data, dict) else None
+
+
+def iter_installed_books(root: Path | None = None) -> list[str]:
+    root_p = books_root(root)
+    if not root_p.is_dir():
+        return []
+    out: list[str] = []
+    for child in sorted(root_p.iterdir()):
+        if not child.is_dir():
+            continue
+        if (child / "search.jsonl").is_file():
+            out.append(child.name)
+    return out
+
+
+def civitai_search_to_book_row(row: dict[str, Any], book_id: str = BOOK_CIVITAI) -> dict[str, Any]:
+    prompt = str(row.get("prompt") or "").strip()
+    neg = str(row.get("negative") or row.get("negativePrompt") or "").strip()
+    parts = [prompt] if prompt else []
+    if neg:
+        parts.append(f"Negative: {neg}")
+    params = row.get("params") if isinstance(row.get("params"), dict) else {}
+    for key, label in (("steps", "steps"), ("cfg", "cfg"), ("cfgScale", "cfg"), ("sampler", "sampler")):
+        if params.get(key) is not None:
+            parts.append(f"{label}: {params[key]}")
+    text = "\n".join(parts)[:EXCERPT_MAX]
+    meta = {
+        "prompt": prompt,
+        "negative": neg,
+        "params": params,
+        "loras": list(row.get("loras") or []),
+        "modelVersionId": row.get("modelVersionId"),
+        "kind": row.get("kind"),
+        "score": row.get("score"),
+    }
+    rid = row.get("id")
+    return {
+        "id": f"{book_id}:{rid}",
+        "book": book_id,
+        "title": "",
+        "tags": list(row.get("tags") or []),
+        "text": text,
+        "body": prompt[:2000] if prompt else text,
+        "rating": row.get("rating") or "pg",
+        "meta": meta,
+    }
+
+
+def _chunk_text(text: str, *, chunk_chars: int = CHUNK_CHARS, overlap: int = CHUNK_OVERLAP) -> list[str]:
+    text = re.sub(r"\r\n?", "\n", text.strip())
+    if not text:
+        return []
+    paras = [p.strip() for p in re.split(r"\n\s*\n", text) if p.strip()]
+    chunks: list[str] = []
+    for para in paras:
+        if len(para) <= chunk_chars:
+            chunks.append(para)
+            continue
+        start = 0
+        while start < len(para):
+            end = min(len(para), start + chunk_chars)
+            chunks.append(para[start:end])
+            if end >= len(para):
+                break
+            start = max(0, end - overlap)
+    return chunks
+
+
+def fictext_row_to_book_rows(
+    row: dict[str, Any],
+    book_id: str,
+    *,
+    row_index: int,
+) -> list[dict[str, Any]]:
+    title = str(row.get("title") or "").strip()
+    tags_raw = row.get("tags")
+    tags: list[str] = []
+    if isinstance(tags_raw, list):
+        tags = [str(t).strip() for t in tags_raw if str(t).strip()]
+    elif isinstance(tags_raw, str):
+        tags = [t.strip() for t in re.split(r"[,;]", tags_raw) if t.strip()]
+    text = str(row.get("text") or "").strip()
+    if not text or looks_minor(text, tags):
+        return []
+    out: list[dict[str, Any]] = []
+    for ci, chunk in enumerate(_chunk_text(text)):
+        excerpt = chunk[:EXCERPT_MAX]
+        out.append(
+            {
+                "id": f"{book_id}:{row_index}:{ci}",
+                "book": book_id,
+                "title": title,
+                "tags": tags[:20],
+                "text": excerpt,
+                "body": chunk[:2000],
+                "rating": "r" if book_id == BOOK_FICTEXT_RPLUS else "pg13",
+                "meta": {"source_row": row_index, "chunk": ci},
+            }
+        )
+    return out
+
+
+def build_civitai_book(*, out_root: Path | None = None, log: Log = _log) -> dict[str, int]:
+    """Write books/civitai-krea2 from datasets/civitai/search.jsonl."""
+    root = out_root or app_root()
+    src = root / "datasets" / "civitai" / "search.jsonl"
+    if not src.is_file():
+        # fallback: build from catalog images via civitai_dataset split output path
+        raise GpuRentError(f"Нет {src} — сначала civitai-dataset split")
+    rows = read_jsonl(src)
+    book_rows = [civitai_search_to_book_row(r) for r in rows]
+    bdir = book_dir(BOOK_CIVITAI, root)
+    write_book_yaml(BOOK_CIVITAI, root)
+    n = write_jsonl(bdir / "search.jsonl", book_rows)
+    write_meta(BOOK_CIVITAI, row_count=n, root=root, extra={"source": str(src)})
+    log(f"books/{BOOK_CIVITAI}: {n} rows → {bdir}")
+    return {"rows": n}
+
+
+def _hf_download_url(filename: str) -> str:
+    return f"https://huggingface.co/datasets/{HF_REPO}/resolve/main/{filename}"
+
+
+def _require_pyarrow():
+    try:
+        import pyarrow.parquet as pq  # noqa: F401
+
+        return pq
+    except ImportError as exc:
+        raise GpuRentError(
+            "Нужен pyarrow для fictext: pip install 'gpu-rent[books]' или pip install pyarrow"
+        ) from exc
+
+
+def download_fictext_parquet(
+    book_id: str,
+    *,
+    out_root: Path | None = None,
+    log: Log = _log,
+    token: str | None = None,
+) -> Path:
+    if book_id not in HF_FILES:
+        raise GpuRentError(f"Unknown fictext book: {book_id}")
+    filename = HF_FILES[book_id]
+    bdir = book_dir(book_id, out_root)
+    bdir.mkdir(parents=True, exist_ok=True)
+    dest = bdir / filename
+    url = _hf_download_url(filename)
+    headers: dict[str, str] = {}
+    if token:
+        headers["Authorization"] = f"Bearer {token}"
+    log(f"download: {url}")
+    with httpx.Client(follow_redirects=True, timeout=600.0) as client:
+        with client.stream("GET", url, headers=headers) as resp:
+            resp.raise_for_status()
+            with dest.open("wb") as fh:
+                for chunk in resp.iter_bytes(1024 * 1024):
+                    fh.write(chunk)
+    log(f"saved {dest} ({dest.stat().st_size // 1024} KB)")
+    return dest
+
+
+def build_fictext_book(
+    book_id: str,
+    *,
+    out_root: Path | None = None,
+    log: Log = _log,
+    max_rows: int | None = None,
+) -> dict[str, int]:
+    if book_id not in HF_FILES:
+        raise GpuRentError(f"Unknown fictext book: {book_id}")
+    pq = _require_pyarrow()
+    bdir = book_dir(book_id, out_root)
+    parquet = bdir / HF_FILES[book_id]
+    if not parquet.is_file():
+        raise GpuRentError(f"Нет {parquet} — сначала books download fictext")
+    write_book_yaml(book_id, out_root)
+    table = pq.read_table(parquet)
+    data = table.to_pydict()
+    titles = data.get("title") or []
+    tags_col = data.get("tags") or []
+    texts = data.get("text") or []
+    n_src = len(texts)
+    book_rows: list[dict] = []
+    for i in range(n_src):
+        if max_rows is not None and i >= max_rows:
+            break
+        row = {
+            "title": titles[i] if i < len(titles) else "",
+            "tags": tags_col[i] if i < len(tags_col) else [],
+            "text": texts[i] if i < len(texts) else "",
+        }
+        book_rows.extend(fictext_row_to_book_rows(row, book_id, row_index=i))
+    n = write_jsonl(bdir / "search.jsonl", book_rows)
+    write_meta(
+        book_id,
+        row_count=n,
+        root=out_root,
+        source_sha=sha256_file(parquet),
+        extra={"source_parquet": parquet.name, "source_stories": n_src},
+    )
+    log(f"books/{book_id}: {n} chunks from {n_src} stories → {bdir}")
+    return {"rows": n, "stories": n_src}
+
+
+def cmd_download_fictext(
+    *,
+    variant: str = "both",
+    out_root: Path | None = None,
+    log: Log = _log,
+) -> None:
+    cfg = load_config()
+    token = (getattr(cfg, "hf_token", None) or "").strip() or None
+    ids = []
+    if variant in {"rplus", "both"}:
+        ids.append(BOOK_FICTEXT_RPLUS)
+    if variant in {"full", "both"}:
+        ids.append(BOOK_FICTEXT_FULL)
+    for book_id in ids:
+        download_fictext_parquet(book_id, out_root=out_root, log=log, token=token)
+        build_fictext_book(book_id, out_root=out_root, log=log)
+
+
+def remote_book_dir(book_id: str) -> str:
+    return f"{VM_BOOKS_PREFIX}/{book_id}"
+
+
+def seed_book_files(local_book: Path, remote_book: str) -> list[tuple[Path, str]]:
+    """Local book dir files to push."""
+    pairs: list[tuple[Path, str]] = []
+    for name in ("book.yaml", "search.jsonl", "meta.json"):
+        local = local_book / name
+        if local.is_file():
+            pairs.append((local, f"{remote_book}/{name}"))
+    meta = local_book / "meta.json"
+    if meta.is_file():
+        pairs.append((meta, f"{remote_book}/{SIDEcar_NAME}"))
+    return pairs
+
+
+def build_parser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(
+        prog="python -m gpu_rent.books",
+        description="Build Assistent reference books (search.jsonl + book.yaml)",
+    )
+    p.add_argument("--out", type=Path, default=None, help="App root")
+    sub = p.add_subparsers(dest="cmd")
+
+    d = sub.add_parser("download", help="Download HF fictext parquet(s)")
+    d.add_argument(
+        "target",
+        choices=["fictext"],
+        nargs="?",
+        default="fictext",
+    )
+    d.add_argument("--variant", choices=["rplus", "full", "both"], default="both")
+
+    b = sub.add_parser("build", help="Build book search.jsonl")
+    b.add_argument("target", choices=["civitai", "fictext", "all"])
+    b.add_argument("--variant", choices=["rplus", "full", "both"], default="both")
+
+    sub.add_parser("list", help="List local books with meta")
+    return p
+
+
+def main(argv: list[str] | None = None) -> int:
+    argv = list(sys.argv[1:] if argv is None else argv)
+    parser = build_parser()
+    if not argv:
+        argv = ["list"]
+    args = parser.parse_args(argv)
+    out = args.out
+    try:
+        if args.cmd == "download":
+            if args.target == "fictext":
+                cmd_download_fictext(variant=args.variant, out_root=out)
+        elif args.cmd == "build":
+            if args.target in {"civitai", "all"}:
+                build_civitai_book(out_root=out)
+            if args.target in {"fictext", "all"}:
+                for book_id in (
+                    [BOOK_FICTEXT_RPLUS, BOOK_FICTEXT_FULL]
+                    if args.variant == "both"
+                    else [BOOK_FICTEXT_RPLUS if args.variant == "rplus" else BOOK_FICTEXT_FULL]
+                ):
+                    build_fictext_book(book_id, out_root=out)
+        elif args.cmd == "list":
+            for bid in iter_installed_books(out):
+                meta = read_meta(bid, out) or {}
+                log(
+                    f"{bid}: rows={meta.get('row_count', '?')} "
+                    f"sha={str(meta.get('content_sha', ''))[:12]}"
+                )
+        else:
+            parser.print_help()
+            return 0
+    except GpuRentError as exc:
+        print(f"error: {exc}", file=sys.stderr)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
@@ -1,4 +1,4 @@
-"""Local Civitai Krea2 gallery scrape → train.jsonl + search.jsonl (no image files)."""
+"""Local Civitai Krea2 gallery scrape → search.jsonl + Assistent book (no image files)."""

 from __future__ import annotations

@@ -522,28 +522,6 @@ def cmd_scrape(
    return have


-def train_row(row: dict[str, Any]) -> dict[str, Any]:
-    tags = row.get("tags") or []
-    tag_s = ", ".join(str(t) for t in tags[:12])
-    rating = row.get("rating") or "pg"
-    instruction = f"Write a Krea 2 prompt.\nTags: {tag_s}\nRating: {rating}"
-    params = row.get("params") if isinstance(row.get("params"), dict) else {}
-    parts = [str(row.get("prompt") or "").strip()]
-    neg = str(row.get("negativePrompt") or "").strip()
-    if neg:
-        parts.append(f"Negative: {neg}")
-    for key, label in (
-        ("steps", "steps"),
-        ("cfgScale", "cfg"),
-        ("sampler", "sampler"),
-        ("seed", "seed"),
-        ("size", "size"),
-    ):
-        if params.get(key) is not None:
-            parts.append(f"{label}: {params[key]}")
-    return {"instruction": instruction, "output": "\n".join(parts)}
-
-
 def search_row(row: dict[str, Any]) -> dict[str, Any]:
    resources = row.get("resources") if isinstance(row.get("resources"), list) else []
    loras: list[dict[str, Any]] = []
@@ -600,14 +578,12 @@ def cmd_split(*, out_root: Path | None = None, log: Log = _log) -> dict[str, int

    by_kind: dict[str, list] = {"checkpoint": [], "lora": []}
    by_rating: dict[str, list] = {}
-    train: list[dict] = []
    search: list[dict] = []
    for row in rows:
        kind = str(row.get("kind") or "checkpoint")
        by_kind.setdefault(kind, []).append(row)
        rating = str(row.get("rating") or "pg")
        by_rating.setdefault(rating, []).append(row)
-        train.append(train_row(row))
        search.append(search_row(row))

    counts: dict[str, int] = {}
@@ -618,13 +594,19 @@ def cmd_split(*, out_root: Path | None = None, log: Log = _log) -> dict[str, int
        safe = re.sub(r"[^a-z0-9]+", "", rating.lower()) or "pg"
        path = root / "by_rating" / f"{safe}.jsonl"
        counts[f"rating:{safe}"] = write_jsonl(path, items)
-    counts["train"] = write_jsonl(root / "train.jsonl", train)
    counts["search"] = write_jsonl(root / "search.jsonl", search)
    log(
        "split: "
        + ", ".join(f"{k}={v}" for k, v in sorted(counts.items()))
        + f" → {root}"
    )
+    try:
+        from gpu_rent.books import build_civitai_book
+
+        book_counts = build_civitai_book(out_root=out_root or app_root(), log=log)
+        counts["book:civitai-krea2"] = book_counts.get("rows", 0)
+    except GpuRentError as exc:
+        log(f"book build warn: {exc}")
    return counts


@@ -657,7 +639,7 @@ def cmd_all(
 def build_parser() -> argparse.ArgumentParser:
    p = argparse.ArgumentParser(
        prog="python -m gpu_rent.civitai_dataset",
-        description="Scrape Civitai Krea2 galleries → train.jsonl + search.jsonl",
+        description="Scrape Civitai Krea2 galleries → search.jsonl + books/civitai-krea2",
    )
    p.add_argument(
        "--out",
@@ -676,7 +658,7 @@ def build_parser() -> argparse.ArgumentParser:
    s.add_argument("--per-version", type=int, default=DEFAULT_PER_VERSION)
    s.add_argument("--min-score", type=int, default=DEFAULT_MIN_SCORE)

-    sub.add_parser("split", help="Write by_kind / by_rating / train / search")
+    sub.add_parser("split", help="Write by_kind / by_rating / search + civitai book")

    a = sub.add_parser("all", help="discover + scrape + split (default)")
    a.add_argument("--top-checkpoints", type=int, default=DEFAULT_TOP_CHECKPOINTS)
@@ -1007,6 +1007,20 @@ def seed_extensions_cmd() -> None:
        _die(exc)


+@app.command("seed-books")
+def seed_books_cmd(
+    force: bool = typer.Option(False, "--force", help="Push all books even if sha unchanged"),
+) -> None:
+    """Push books/*/ → VM Assistent/books/ (changed only unless --force)."""
+    try:
+        from gpu_rent.provision import seed_books
+
+        cfg, host = _live()
+        seed_books(cfg, host, log, force=force)
+    except GpuRentError as exc:
+        _die(exc)
+
+
@app.command("seed-personas")
 def seed_personas_cmd() -> None:
    """Push assistent-extensions/ → VM Assistent/extensions (+ _base)."""
@@ -525,6 +525,16 @@ def build_chat_trace(
    }
    if isinstance(response, dict) and response.get("civitai_results"):
        out["civitai_results"] = response.get("civitai_results")
+    if isinstance(response, dict) and isinstance(response.get("knowledge"), dict):
+        out["knowledge"] = response.get("knowledge")
+    elif isinstance(response, dict) and response.get("civitai_results"):
+        # Shim until swarm-assistent ships unified knowledge block
+        out["knowledge"] = {
+            "catalog": [],
+            "hops": [],
+            "results": response.get("civitai_results"),
+            "legacy": "civitai_results",
+        }
    return out


@@ -72,6 +72,14 @@ def assistent_personas_manifest_path() -> Path:
    return app_root() / "assistent-personas.yaml"


+def books_dir() -> Path:
+    return app_root() / "books"
+
+
+def books_example_dir() -> Path:
+    return app_root() / "books.example"
+
+
 def assistent_personas_example_path() -> Path:
    """Deprecated yaml example — prefer assistent_extensions_example_dir()."""
    return app_root() / "assistent-personas.example.yaml"
@@ -1141,11 +1141,81 @@ def seed_assistent_personas(cfg: Config, host: str, log: Log) -> None:
        f"assistent-extensions → packs/{pushed} "
        f"(default={default_id}{ctx_note})"
    )
+    seed_books(cfg, host, log)
+
+
+def seed_books(cfg: Config, host: str, log: Log, *, force: bool = False) -> None:
+    """Push changed books/*/ → Assistent/books/ (content_sha diff)."""
+    import json
+
+    from gpu_rent.books import (
+        SIDEcar_NAME,
+        VM_BOOKS_PREFIX,
+        iter_installed_books,
+        read_meta,
+        remote_book_dir,
+        seed_book_files,
+    )
+    from gpu_rent.paths import app_root
+    from gpu_rent.ssh_ops import put_file, put_text
+
+    root = Path(getattr(cfg, "app_root", None) or app_root())
+    book_ids = iter_installed_books(root)
+    if not book_ids:
        seed_civitai_examples(cfg, host, log)
+        return
+
+    run_ssh(cfg, host, f"mkdir -p {shlex.quote(VM_BOOKS_PREFIX)}", check=False)
+    pushed = 0
+    for book_id in book_ids:
+        meta = read_meta(book_id, root)
+        local_sha = str((meta or {}).get("content_sha") or "").strip()
+        if not local_sha:
+            log(f"books/{book_id}: нет meta.content_sha — skip")
+            continue
+        remote_dir = remote_book_dir(book_id)
+        sidecar_remote = f"{remote_dir}/{SIDEcar_NAME}"
+        remote_raw = run_ssh(
+            cfg,
+            host,
+            f"test -f {shlex.quote(sidecar_remote)} && cat {shlex.quote(sidecar_remote)} || true",
+            check=False,
+            timeout=20,
+        ).strip()
+        remote_sha = ""
+        if remote_raw:
+            try:
+                remote_meta = json.loads(remote_raw)
+                if isinstance(remote_meta, dict):
+                    remote_sha = str(remote_meta.get("content_sha") or "")
+            except json.JSONDecodeError:
+                remote_sha = ""
+
+        if not force and remote_sha == local_sha:
+            log(f"books/{book_id}: unchanged")
+            continue
+
+        run_ssh(cfg, host, f"mkdir -p {shlex.quote(remote_dir)}", check=False)
+        local_dir = root / "books" / book_id
+        for local_path, remote_path in seed_book_files(local_dir, remote_dir):
+            put_file(cfg, host, local_path, remote_path)
+        sidecar_payload = json.dumps(
+            {"content_sha": local_sha, "book_id": book_id, **(meta or {})},
+            ensure_ascii=False,
+        )
+        put_text(cfg, host, sidecar_remote, sidecar_payload + "\n")
+        size_kb = max(1, (local_dir / "search.jsonl").stat().st_size // 1024)
+        log(f"books/{book_id} -> {remote_dir} ({size_kb} KB, sha={local_sha[:12]})")
+        pushed += 1
+
+    if pushed:
+        log(f"books: pushed {pushed}/{len(book_ids)}")
+    elif book_ids:
+        log(f"books: all {len(book_ids)} unchanged")


 def seed_civitai_examples(cfg: Config, host: str, log: Log) -> None:
-    """Push datasets/civitai/search.jsonl → Assistent/civitai-examples.jsonl (FTS, no embed)."""
+    """Legacy FTS path when books/ not built yet."""
    from gpu_rent.paths import app_root
    from gpu_rent.ssh_ops import put_file

@@ -56,7 +56,7 @@ def test_seed_assistent_personas_pushes_packs(tmp_path: Path, monkeypatch):
        "gpu_rent.paths.assistent_personas_manifest_path", lambda: tmp_path / "no.yaml"
    )
    monkeypatch.setattr(
-        "gpu_rent.provision.seed_civitai_examples", lambda *a, **k: None
+        "gpu_rent.provision.seed_books", lambda *a, **k: None
    )

    cfg = MagicMock()
@@ -98,7 +98,7 @@ def test_seed_skips_empty_pack_dir(tmp_path: Path, monkeypatch):
        "gpu_rent.paths.assistent_personas_manifest_path", lambda: tmp_path / "no.yaml"
    )
    monkeypatch.setattr(
-        "gpu_rent.provision.seed_civitai_examples", lambda *a, **k: None
+        "gpu_rent.provision.seed_books", lambda *a, **k: None
    )

    cfg = MagicMock()
@@ -0,0 +1,129 @@
+"""Tests for Assistent books build + seed helpers."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from unittest.mock import MagicMock
+
+from gpu_rent.books import (
+    BOOK_CIVITAI,
+    build_civitai_book,
+    civitai_search_to_book_row,
+    fictext_row_to_book_rows,
+    read_meta,
+    sha256_file,
+    write_meta,
+)
+from gpu_rent.provision import seed_books
+
+
+def test_civitai_search_to_book_row():
+    row = {
+        "id": 1,
+        "prompt": "woman, redhead",
+        "negative": "blur",
+        "tags": ["woman"],
+        "rating": "pg13",
+        "params": {"steps": 8, "cfg": 1},
+        "loras": [{"versionId": 9, "weight": 0.7}],
+        "modelVersionId": 100,
+        "kind": "checkpoint",
+        "score": 10,
+    }
+    out = civitai_search_to_book_row(row)
+    assert out["book"] == BOOK_CIVITAI
+    assert "woman" in out["text"]
+    assert out["meta"]["loras"][0]["versionId"] == 9
+
+
+def test_fictext_chunks_skip_minor():
+    rows = fictext_row_to_book_rows(
+        {"title": "T", "tags": ["loli"], "text": "loli content here"},
+        "ru-fictext-rplus",
+        row_index=0,
+    )
+    assert rows == []
+
+
+def test_fictext_chunks_ok():
+    text = "First paragraph.\n\nSecond paragraph with more words."
+    rows = fictext_row_to_book_rows(
+        {"title": "Story", "tags": ["romance"], "text": text},
+        "ru-fictext-rplus",
+        row_index=3,
+    )
+    assert len(rows) >= 1
+    assert rows[0]["title"] == "Story"
+    assert "romance" in rows[0]["tags"]
+
+
+def test_build_civitai_book(tmp_path: Path):
+    ds = tmp_path / "datasets" / "civitai"
+    ds.mkdir(parents=True)
+    search_row = {
+        "id": 5,
+        "prompt": "test prompt",
+        "negative": "",
+        "tags": ["a"],
+        "rating": "pg",
+        "params": {},
+        "loras": [],
+    }
+    (ds / "search.jsonl").write_text(json.dumps(search_row) + "\n", encoding="utf-8")
+    counts = build_civitai_book(out_root=tmp_path, log=lambda m: None)
+    assert counts["rows"] == 1
+    meta = read_meta(BOOK_CIVITAI, tmp_path)
+    assert meta and meta.get("row_count") == 1
+    book_search = tmp_path / "books" / BOOK_CIVITAI / "search.jsonl"
+    assert book_search.is_file()
+    assert sha256_file(book_search) == meta["content_sha"]
+
+
+def test_seed_books_skips_unchanged(tmp_path: Path, monkeypatch):
+    book = tmp_path / "books" / BOOK_CIVITAI
+    book.mkdir(parents=True)
+    (book / "search.jsonl").write_text('{"id":"x"}\n', encoding="utf-8")
+    write_meta(BOOK_CIVITAI, row_count=1, root=tmp_path)
+    meta = read_meta(BOOK_CIVITAI, tmp_path)
+    sidecar = json.dumps({"content_sha": meta["content_sha"]})
+
+    monkeypatch.setattr("gpu_rent.paths.app_root", lambda: tmp_path)
+    monkeypatch.setattr(
+        "gpu_rent.provision.run_ssh",
+        lambda *a, **k: sidecar if ".gpu-rent-meta" in str(a[2]) else "",
+    )
+    put_calls: list[str] = []
+    monkeypatch.setattr(
+        "gpu_rent.ssh_ops.put_file",
+        lambda *a, **k: put_calls.append(str(a[3])),
+    )
+    monkeypatch.setattr("gpu_rent.ssh_ops.put_text", lambda *a, **k: None)
+    monkeypatch.setattr("gpu_rent.provision.seed_civitai_examples", lambda *a, **k: None)
+
+    logs: list[str] = []
+    seed_books(MagicMock(app_root=str(tmp_path)), "host", logs.append)
+    assert put_calls == []
+    assert any("unchanged" in m for m in logs)
+
+
+def test_seed_books_pushes_when_sha_diff(tmp_path: Path, monkeypatch):
+    book = tmp_path / "books" / BOOK_CIVITAI
+    book.mkdir(parents=True)
+    (book / "book.yaml").write_text("kind: book\n", encoding="utf-8")
+    (book / "search.jsonl").write_text('{"id":"x"}\n', encoding="utf-8")
+    write_meta(BOOK_CIVITAI, row_count=1, root=tmp_path)
+
+    monkeypatch.setattr("gpu_rent.paths.app_root", lambda: tmp_path)
+    monkeypatch.setattr("gpu_rent.provision.run_ssh", lambda *a, **k: "")
+    put_calls: list[str] = []
+    monkeypatch.setattr(
+        "gpu_rent.ssh_ops.put_file",
+        lambda *a, **k: put_calls.append(str(a[3])),
+    )
+    monkeypatch.setattr("gpu_rent.ssh_ops.put_text", lambda *a, **k: None)
+
+    logs: list[str] = []
+    seed_books(MagicMock(app_root=str(tmp_path)), "host", logs.append)
+    assert any("search.jsonl" in p for p in put_calls)
+    assert any("books/civitai-krea2" in m for m in logs)
@@ -18,7 +18,6 @@ from gpu_rent.civitai_dataset import (
    reaction_score,
    search_row,
    tags_from_image,
-    train_row,
 )
 from gpu_rent.errors import CloudError

@@ -108,7 +107,7 @@ def test_normalize_skips_low_score():
    )


-def test_train_and_search_rows():
+def test_search_row():
    row = {
        "id": 7,
        "rating": "r",
@@ -121,11 +120,6 @@ def test_train_and_search_rows():
        "params": {"steps": 8, "cfgScale": 1, "sampler": "euler"},
        "resources": [{"type": "lora", "modelVersionId": 1, "weight": 0.5}],
    }
-    tr = train_row(row)
-    assert "Tags: woman, cinematic" in tr["instruction"]
-    assert "Rating: r" in tr["instruction"]
-    assert "A woman" in tr["output"]
-    assert "cfg: 1" in tr["output"]
    sr = search_row(row)
    assert sr["id"] == 7
    assert sr["loras"] == [{"versionId": 1, "weight": 0.5}]
@@ -165,16 +159,16 @@ def test_cmd_split_writes_artifacts(tmp_path: Path):
        for r in rows:
            fh.write(json.dumps(r) + "\n")
    counts = cmd_split(out_root=tmp_path, log=lambda m: None)
-    assert counts["train"] == 2
    assert counts["search"] == 2
    assert counts["kind:checkpoint"] == 1
    assert counts["kind:lora"] == 1
    assert counts["rating:pg"] == 1
    assert counts["rating:x"] == 1
+    assert counts.get("book:civitai-krea2") == 2
    search = (tmp_path / "datasets" / "civitai" / "search.jsonl").read_text(encoding="utf-8")
    assert '"id": 1' in search
-    train = (tmp_path / "datasets" / "civitai" / "train.jsonl").read_text(encoding="utf-8")
-    assert "Write a Krea 2 prompt" in train
+    book = tmp_path / "books" / "civitai-krea2" / "search.jsonl"
+    assert book.is_file()


 def _mock_response(payload: dict, status: int = 200) -> httpx.Response:
@@ -2,7 +2,7 @@
 sudo -n systemctl restart swarmui
 sleep 40
 S=$(curl -s -X POST http://127.0.0.1:7801/API/GetNewSession -H 'Content-Type: application/json' -d '{}' | sed -n 's/.*"session_id": *"\([^"]*\)".*/\1/p')
-for api in AssistentListChats AssistentListTrainSamples AssistentListMemory; do
+for api in AssistentListChats AssistentListMemory; do
  echo "--- $api ---"
  curl -s -X POST "http://127.0.0.1:7801/API/$api" -H 'Content-Type: application/json' -d "{\"session_id\":\"$S\"}" | head -c 300
  echo